Volt
An open-source workspace for working with coding agents.
Chats, code, Git, a browser and a terminal in one window, on top of the real VS Code core.
Download · Website · Docs · Changelog · Contributing · Security
Agents do the typing. You keep the judgment. Volt is a workspace for working with coding agents instead of next to them: the chat, the code, Git, a browser and a terminal share one window, so nothing you need is a tab away.
It is built on the real VS Code core, not a lookalike, so your extensions, themes, keybindings, language servers and debuggers come along. What Volt adds is the layer that makes agents a first-class part of the editor: a runtime that talks to any model or agent CLI, a permission system you control, and review tools that let you keep or undo exactly what an agent changed.
Public beta. Volt is below 1.0 and moves fast, so expect rough edges and changes between releases. The changelog lists what changed in each one.
Download
Get the latest build from GitHub Releases.
| Platform | File |
|---|
| macOS, Apple silicon | volt-stable-<version>-darwin-arm64.dmg |
| macOS, Intel | volt-stable-<version>-darwin-x64.dmg |
| Windows | volt-stable-<version>-win32-<x64 or arm64>-user-setup.exe |
| Linux | volt-stable-<version>-linux-<x64, arm64 or armhf> as .deb, .rpm or .tar.gz |
The builds are not code-signed yet, because Volt is an independent project without a signing certificate. Your system warns you the first time you open one. That is expected:
- macOS: open System Settings > Privacy & Security and click Open Anyway, or install from the Terminal and skip the dialogs. The macOS install guide walks through it with a screenshot for every step, and shows how to use Volt from the command line (
volt .).
- Windows: when SmartScreen says Windows protected your PC, click More info, then Run anyway.
- Linux:
sudo apt install ./volt-stable-<version>-linux-x64.deb, or sudo dnf install ./volt-stable-<version>-linux-x64.rpm.
There is also a Beta and a Nightly channel, each its own app that installs next to Volt (see Releasing Volt). macOS is the platform Volt is developed and tested on; Windows and Linux builds come from the same pipeline but have had little testing. Everything about installing is in docs/install.
Why Volt
| Principle | What it means |
|---|
| Bring your own agent | Volt doesn't sell tokens. Use the Claude Code, Codex, OpenCode, Cursor, Grok, Antigravity or Kimi sign-in you already have, or point Volt at an API (OpenAI, Anthropic, Gemini, OpenRouter) or a local model (Ollama, LM Studio). You keep your plan and your keys. |
| You set the leash | Four access presets, from Supervised to Full access, decide what an agent may read, edit, run, fetch or connect to. Risky commands are flagged before they run, and a hard-deny list for destructive commands and secret files holds even under Full access. |
| Review before it lands | Every agent edit is pending until you keep or undo it, hunk by hunk. Each turn is checkpointed in Git, so you can preview, restore or redo any point in a conversation without touching your own index. |
| One window, no tab sprawl | An in-app browser the agent can drive and you can annotate, a real editor, Git and a terminal sit beside the conversation. Switch between the Agent and IDE layouts with one command. |
| Open all the way down | MIT licensed, with extensions from Open VSX. If we ever go the wrong way, everything you need to fork it is here. |
What's inside
Agents and models
- Agent CLIs over ACP. Claude Code, Codex, OpenCode, Cursor (early access), Grok (early access), Antigravity, Kimi and any generic Agent Client Protocol agent. Warm sessions are kept ready, so a new chat doesn't pay a cold start.
- Direct model access. OpenAI, Anthropic, Gemini, OpenRouter, any OpenAI-compatible endpoint, Ollama and LM Studio, driven by Volt's own native tool loop.
- Five modes in one composer. Agent, Plan, Ask, Debug and Multitask. Plan and Ask can't edit files or run shell commands. Plan mode proposes a plan card you approve, revise, edit or save.
- Subagents. Delegate work to subagents, watch them in the side toolbar, read their chats, and resume a stopped one or retry a failed one. Each subagent's changes show up in the turn's review.
- Several Claude accounts. Sign in to more than one login and switch by hand, or let Volt move to another account near a usage limit.
- MCP. Connect MCP servers (stdio or HTTP) from
.mcp.json, .cursor/mcp.json, .vscode/mcp.json or .volt/mcp.json. Agents call them like built-in tools, under the same permissions.
- Project instructions.
AGENTS.md, CLAUDE.md, .volt/AGENTS.md, .cursorrules and .github/copilot-instructions.md are picked up automatically.
- A code index for agents. Volt indexes your open projects in the background, so
grep, glob and symbol lookups answer from an index instead of walking the tree.
Control and review
- Access presets. Supervised (default), Auto-accept edits, Auto and Full access, layered with project, agent, session and mode rules. Allow-always decisions are remembered and visible.
- Pending edits. Keep or undo changes hunk by hunk against a baseline, with Undo All and Keep All a click away in the Changes header.
- Checkpoints. A Git snapshot before and after every turn, stored under a hidden ref (
refs/volt/*), with preview, restore and redo. Your index and branches stay untouched.
- Per-chat worktrees and sandboxes. Give a chat its own checkout so parallel agents don't trample each other, and choose its sandbox level and network.
The workspace around the chat
- Agent and IDE layouts. One command toggles between a chat-first window and the full editor workbench.
- Projects. Add a project from this computer, a Git URL or GitHub, and switch between projects and chats from one sidebar.
- In-app browser with design mode. The agent can navigate, click, type, take screenshots and read the console and network. You can pick an element or draw an area, name what to change, and batch several annotations into one message.
- Git, built in. Branches, Remotes, Tags, Worktrees, Stashes and Commits sit in Source Control, and pull requests open in Volt (GitHub, with reading support for GitLab, Bitbucket, Gitea and Azure DevOps).
- A rich composer.
@-mention files, folders, terminals, other chats, branches, editor selections, browser elements and images. Queue your next message while the agent works, or dictate it.
- Questions, plans and todos. Agents can ask you a structured question, propose a plan you approve before it runs, and show their task list as they go.
- Context and usage meters. See how much of a model's context a chat is using, and what Claude, Codex and Cursor usage costs, from local transcripts and account data.
In the editor
- Tab completion and next-edit prediction. Inline suggestions from the model you choose, plus a one-shot Volt: AI Edit command for multi-file changes. Disabled for
.env*, lockfiles and secrets by default.
- Everything VS Code. Debugging, tasks, the extension host, remote development and the full set of built-in language features.
How it fits together
The workbench never embeds a provider SDK. The UI talks to a runtime service; the runtime decides which model or agent runs, asks the access broker before any tool touches your machine, and uses platform services for processes, Git and the loopback MCP server that exposes Volt's browser and question tools to agents.
| Path | What it is |
|---|
src/vs/workbench/services/voltRuntime | The runtime: providers, ACP bridges, the native loop, the access broker, tools and history |
src/vs/workbench/contrib/voltAgent | The agent window: chat, composer, review, in-app browser and history |
src/vs/workbench/contrib/voltProjects | The project registry and the add-project flow |
src/vs/workbench/contrib/voltSettings | The Volt settings editor |
src/vs/workbench/contrib/voltPrediction | Inline completions and one-shot AI edits |
src/vs/platform /voltStdio, /voltGit, /voltHostMcp, /voltUsage, /voltFsBrowse, /voltBrowser | Host services for processes, Git snapshots, MCP, usage, folder browsing and the browser session |
apps/docs | The documentation site and landing page |
brand | Icon sources and generated assets |
Everything else is the VS Code core, kept as close to upstream as we can so that merging new releases stays tractable. VOLT-ARCHITECTURE-REVIEW.md is a candid engineering review, and VOLT-MULTI-PROJECT-AGENT-WORKSPACE-PLAN.md describes where the agent window is headed.
Getting started
Connect an agent
Open Volt Settings from the command palette (⇧⌘P) and choose how you want to run agents:
- An agent CLI you already use. Install it and sign in the way that tool documents. Volt finds it on your
PATH (or launches a pinned adapter through npx) and leaves your plan and credentials with that tool. Volt does not store them.
- An API key. Add a provider (OpenAI, Anthropic, Gemini, OpenRouter or any OpenAI-compatible endpoint). Keys go in the editor's secret storage, not in settings files.
- A local model. Point Volt at Ollama (
127.0.0.1:11434 by default) or LM Studio.
Then press ⌘N for a new agent chat.
Build from source
You need Node.js 22.19 (the version in .nvmrc; 22.15.1 is the minimum), npm, Git, Python 3 and a C/C++ toolchain for native modules. On macOS that is the Xcode Command Line Tools. CONTRIBUTING.md lists the Linux and Windows packages.
git clone https://github.com/LeulAria/volt.git
cd volt
nvm install # picks up Node from .nvmrc
npm install # installs dependencies, builds native modules
npm run electron # downloads the Electron runtime into .build/
node build/lib/builtInExtensions.js # fetches the pinned built-in extensions
make start # compile, watch and launch Volt
make start runs a file watcher and launches the app with hot reload: save a file and the window reloads. make help lists the rest (make stop, make reload, make restart). The first compile takes a few minutes; after that, rebuilds are incremental.
To package a macOS app yourself:
unset ELECTRON_RUN_AS_NODE
VOLT_BUILD_IGNORE_TYPE_ERRORS=1 \
node --max-old-space-size=16384 ./node_modules/gulp/bin/gulp.js vscode-darwin-arm64-min
The unsigned app lands in ../VSCode-darwin-arm64/VOLT.app. VOLT_BUILD_IGNORE_TYPE_ERRORS=1 reports type errors but still packages, because the tree can carry known TypeScript diagnostics; see CONTRIBUTING.md. Release builds for every platform are made by GitHub Actions, as described in Releasing Volt.
Privacy and safety
Volt ships with no telemetry endpoint. What leaves your machine is what you choose to send: prompts, file contents and tool results go to whichever model or agent you select, and tab prediction (on by default) sends the code around your cursor to the model you picked for it. Volt also makes requests to GitHub (when you add a project from GitHub and, in release builds, to check a static update feed for new versions), Open VSX (extensions) and the usage endpoints of the agents you've signed into. Development builds never check for updates.
Permissions are an approval policy, not a sandbox: an agent you allow to run shell commands runs them as you. Read SECURITY.md for the full trust model before you point Volt at code you don't trust.
Documentation
Contributing
Volt is built in the open and we'd love your help. Bug reports with clear reproduction steps, focused fixes and well-tested improvements are the fastest way to get something merged. For a larger feature, open an issue first so we can agree on the shape before you invest the time.
Read CONTRIBUTING.md and the Code of Conduct to get started. To report a security problem, don't open a public issue; follow SECURITY.md.
Acknowledgements
Volt stands on a lot of open work:
Volt is an independent project. It is not affiliated with, endorsed by or sponsored by Microsoft, Anthropic, OpenAI, Google, Cursor or any other company named here. "Visual Studio Code" is a trademark of Microsoft Corporation, and all other product names and logos belong to their owners.
License
Volt is released under the MIT License.
Copyright (c) 2026 - present Volt ADK. Portions copyright (c) 2015 - present Microsoft Corporation.
